2021 Volume 35 Issue 6
Published: 08 July 2021
-
Select all|
-
Survey2021, 35(6): 1-15.Evaluation in natural language processing drives and promotes research on models and methods. In recent years, new evaluation data sets and evaluation tasks have been continuously proposed. At the same time, a series of problems exposed by such evaluations seems to restrict the progress of natural language processing technology. Starting from the concept, composition, development and significance of natural language Processing evaluation, this article classifies and summarizes the tasks and characteristics of mainstream natural language Processing evaluation, and then reveals the problems and their possible causes. In parallel to the human language ability evaluation standard, this paper puts forward the concept of human-like machine language ability evaluation, and proposes a series of basic principles and implementation ideas for human-like machine language ability evaluation from three aspects: reliability, difficulty and validity.
-
Survey2021, 35(6): 16-29.Retweeting, a simple and powerful function, is the key mechanism of information diffusion on major microblog platforms. Retweeting behavior can help us understand the characteristics of information diffusion and better explore users' behaviors and interests, which is of significance in various applications such as information recommendation, emergency prevention and public opinion monitoring. This paper discusses various research work related to predicting retweeting behavior and retweet count, and summarizes the current challenges and future research direction. This paper reviews the research in the field of retweeting so as to provide a reference for future researchers.
-
Language Resources Construction2021, 35(6): 30-38.Target-level sentiment classification task is to get the sentiment tendency of a specific evaluation target in a sentence. There are often multiple targets in a comment sentence, and the sentiments of multiple targets may be consistent or inconsistent. However, in the existing evaluation datasets for target-level sentiment classification: 1) most of them are single sentence with one target; 2) in a few sentences with multiple targets, the sentiment distribution of multiple targets is seriously biased: most multiple targets have the same emotion. In response to the above problems, this paper constructs a Chinese dataset for multi-target sentiment classification, totaling 2,071 items with 6,339 targets manually annotated. The data set provides balance distribution for the number of evaluation targets, positive and negative sentiments, and multi-target sentimental tendency. Meanwhile, this article uses multiple mainstream models of target-level sentiment classification to conduct experiments and comparative analysis on this dataset. Experimental results show that the existing mainstream models are still unable to well classify the targets in instances where there are multiple targets and the target's sentiment is inconsistent, especially when the target's sentiment is neutral.
-
Machine Translation2021, 35(6): 39-46.The Transformer is one of the best performing machine translation models. Generating tokens one by one from left to right, this approach lacks the guidance of future contextual information. To alleviate this issue, we propose a neural machine translation model based on re-decoding. The model treats the generated machine translation outputs as approximate contextual environment of the target language, and then re-decodes each token in the machine translation output successively. The masked multi-head attention of the Transformer decoder only masks the current position token in the generated translation output. As a result, every token re-decoded can make full use of its contextual information. Experimental results on several test sets from the WMT show that the quality of machine translation is improved significantly by leveraging the re-decoding.
-
Machine Translation2021, 35(6): 47-54.As an important task in machine translation, quality estimation of machine translation plays an important role in the development and application of machine translation. In the paper, we propose a simple and effective unified model base on Transformer for quality estimation of machine translation. The model is composed of the Transformer bottleneck layer and a Bi-LSTM network. Parameters of Transformer bottleneck layer are preliminarily optimized with bilingual parallel corpus, and all parameters of the model are jointly optimized and fine-tuned with the training dataset of quality estimation. In model testing, the translation outputs to be estimated are dealt with teacher forcing and a special masking, and then input into the unified model along with the source sentences. The experimental results on the datasets form CWMT18 quality estimation task show that the proposed model is significantly superior to the baseline models trained on the same data, and comparable with that of the best baseline model trained on the large scale bilingual corpus.
-
Information Extraction and Text Mining2021, 35(6): 55-62,73.To deal with model reconstruction process and the lack of training data for various domains in the task of named entity recognition, a domain adaptive named entity recognition method is proposed based on attention mechanism. Firstly, a bidirectional long-short term memory conditional random field named entity recognition model based on the BERT (BERT-BiLSTM-CRF)is constructed on the general dataset. Then, such-bulit model is fine-tuned using the ancient Chinese corpus, with an adaptive neural network layer based on the attention mechanism inserted. The comparison experiment is set with the model in the target domain and the existing transfer learning method. The experimental results show that the proposed model improves the F1 value by 4.31% compared with the generic domain BERT-BiLSTM-CRF model, by 2.46% compared with the same model trained only on the ancient Chinese domain corpus.
-
Information Extraction and Text Mining2021, 35(6): 63-73.Most of the existing named entity recognition methods treat the sentence as a sequence, ignoring the syntactic information in the sentence. This paper proposed a named entity recognition model based on dependency relationship. Adding dependency tree information to the input data, the child and parent node information of words in the dependency tree are obtained by changing the inter layer propagation mode in Bi-LSTM. The features are dynamically selected by the attention mechanism. Finally, the CRF layer is adopted to realize named entity annotation. Experimental results show that the proposed method is better than BiLSTM-CRF model, especially for long entity recognition, achieving 88.94%, 77.42% and 84.38% F1 values on OntoNotes 5.0 English, OntoNotes 5.0 Chinese and semeval-2010 Task 1 Spanish respectively.
-
Information Extraction and Text Mining2021, 35(6): 74-84.Extracting relational triples is a basic task for large-scale knowledge graph construction. In order to improve the ability of extracting overlapped relation triples and multi-slot relation triples, this paper proposes BSLRel, an end-to-end relation triple extraction model based on neural network. Specifically, BSLRel model converts the relation triplet extraction task into a cascade binary sequence labeling task,which consists of a new multiple information fusion structure “Conditional Layer Normalization” to integrate information. With BSLRel, we participate in the “Relation Extraction” task organized by “the 2020 Language and Intelligence Challenge” and achieve Top 5 among all competitive models.
-
Sentiment Analysis and Social Computing2021, 35(6): 85-92.Emotion cause recognition is a new research issue in the field of text emotion analysis. Based on linguistic characteristic of emotion cause, we propose an emotion cause recognition method by emotion context position attention neural network (ECPA). This method considers emotion information such as emotion and category embedding, applies Bi-LSTM to capture context semantic information, and imports a position-based attention mechanism to recognize emotion cause. The experimental results demonstrate the effectiveness of our proposed model, which achieves better performance than state-of-the-art models.
-
Sentiment Analysis and Social Computing2021, 35(6): 93-102.Cross-domain attribute-oriented sentiment analysis is a challenging issue. To explore domain-specific features for cross-domain attribute-oriented sentiment analysis, we divide sentiment words into shared sentiment words and specific sentiment words, and thus propose a domain specific sentiment words attention model (DSSW-ATT). Firstly, we set up two independent subspaces and use attention mechanism to extract shared sentiment word features and specific sentiment word features, respectively. Then, we establish a shared feature classifier and a specific feature classifier. Finally, we use co-training to combine the two kinds of information. To examine our method, we build a couple of attribute-level online review datasets in the hotel domain (as the source domain) and the phone domain (as the target domain). Experimental results show that the proposed method outperforms the baselines.
-
Sentiment Analysis and Social Computing2021, 35(6): 103-111.Irony is a complex linguistic phenomenon that are widely used in social media. With the rapid development Bi-directional of internet, how to recognize irony automatically has become one of the hot topics in the field of natural language processing. this paper proposes an irony recognition based on multiple semantic representation fusion. We use ELMo to train domain vector from large-scale ironic texts, and then combine the semantic representations based on syntactic structure and style information. Bi-directional long short-term memory(Bi-LSTM) network and convolutional neural network are used to recognize irony. Experimental results show that our proposed model can capture the latent semantic features of irony texts from multiple dimensions, bringing a significant improvement on IAC dataset.
-
Sentiment Analysis and Social Computing2021, 35(6): 112-121.It is crucial to understand the relation between social cohesion and the dynamic behavior of social groups. A model of social group dynamics is explored on the basis of cooperative games. Based on the classical network topology structure, this paper investigates the strategy of enhancing cohesion among social groups, and proposes the CPMC and CPIN algorithms which involve the largest clique. Through a specific intervention mechanism, the whole network is divided into two layers, and periphery nodes are selected to join the core layer. Increasing the links between nodes at the same time, the social groups are demonstrated for a better social cohesion.
-
Natural Language Understanding and Generation2021, 35(6): 122-130.Natural language inference is to infer the semantic logical relationship between two given sentences. This paper proposes an inference model to simulate human thinking. Firstly, the context features of sentences are extracted by BiLSTM (bidirectional long short-term memory), which imitates human beings to understand sentence meaning. Then, the semantic graph for every pair of sentences is constructed according to the external semantic knowledge. The spatial features of words are extracted by graph convolutional network or graph attention network, which simulates the thinking mode of analyzing the semantic role similarity of two sentences. Finally, the semantic relationship of two sentences is inferred by integrating the context features and the spatial features. Further analysis reveals that the semantic knowledge is better exploited by graph neural network in natural language inference task.
-
Natural Language Understanding and Generation2021, 35(6): 131-140.Punctuation errors or omissions in s in Chinese texts seriously affects various natural language processing such as semantic analysis and machine translation. Existing researches on punctuation prediction are mostly focused on the speech transcribed text of English conversations, rather than texts in social media and question answering domain. This paper proposes a cross domain Chinese punctuation prediction task, i.e. punctuation prediction for the fields of social media and question answering via large-scale news texts with correct punctuation marks. Corresponding data sets in the fields of news, social media and question answering are then constructed. A BERT-based punctuation prediction baseline model is implemented. The experimental results show that the performance of punctuation prediction in social media and question answering domains decreases by directly using the model trained in the news domain. The decline in question answering domain is much less than that in Weibo domain(more than 20%). The task of cross domain punctuation prediction is challenging.
